Click here to


Are you sure ?

Yes, do it No, cancel

Learning the Plausible VMAT Subspace with Deep Autoencoders

N Bice1*, N Kirby1, R Li1, C Kabat1, D Nguyen2, N Papanikolaou1, M Fakhreddine1, (1) UT Health San Antonio, San Antonio, TX, (2) UT Southwestern Medical Center, Dallas, TX.


(Sunday, 7/12/2020)   [Eastern Time (GMT-4)]

Room: AAPM ePoster Library

Purpose: Volumetric modulated arc therapy (VMAT) planning is a challenging problem in high-dimensional, non-convex optimization. Traditionally, heuristics such as fluence-map-optimization-informed segment initialization use locally optimal solutions to begin the search of the full VMAT plan-space from a reasonable starting point. These greedy routines facilitate VMAT optimization such that clinically satisfactory radiation treatment plans can be created in about 10 minutes. We believe, however, that current VMAT algorithms make compromises in approaching a global optimum and are slower than necessary due to plan overparameterization. In this work, we take a step towards addressing these problems by reducing the effective dimension of VMAT plans with unsupervised deep learning.

Methods: We collected a dataset of 1,874 clinically delivered VMAT arcs from historical treatments documented in Elekta VersaHD log files. We downsampled by gantry angle to 80 evenly-spaced control points and stored multileaf collimator (MLC) leaf positions for 80 leaves on 2 banks (12,800-dimensional parameterization). A family of 2-D fully-convolutional variational autoencoders was used to reduce the dimension of the arc data, and we stored the models that best reconstructed a validation dataset. Given the original and reconstructed apertures, we calculated accumulated fluences for each treatment plan and performed a 2-D gamma analysis to evaluate perturbations to the dose distribution during compression.

Results: The dimension of VMAT MLC data was reduced to 128 with a median position reconstruction error of 4.4 mm. We observed a 2-D gamma pass rate of 72.0% ± 13.3% when comparing the ground truth and reconstructed cumulative fluences on the validation dataset.

Conclusion: Moving forward, we will use transfer learning to improve the model’s performance. We intend to build our deep-learned latent space into an optimization engine, from which we anticipate faster planning times and better conformity to dose constraints.


Inverse Planning, Modeling, Pattern Recognition


TH- External Beam- Photons: IMRT/VMAT dose optimization algorithms

Contact Email